A Fast Robust Human Activity Recognition Method Based on Micro-Doppler Feature Fusion and Broad Learning System

作者：Qun Ma, Zuobin Ying, Qi Zhang, Junwei Duan, Lingzhi Zhu, Weijie Xia, Yi Li · 发表于：IEEE Sensors Journal · 年份：2025 · DOI：10.1109/jsen.2025.3637142 · 被引用次数：1 · 研究领域：Advanced SAR Imaging Techniques、Non-Invasive Vital Sign Monitoring、Sparse and Compressive Sensing Techniques

With the rapid advancement of neural networks and radar sensing, millimeter-wave radar-based human activity recognition (HAR) has become pivotal in sensing applications including robotics, edge-enabled smart homes, and autonomous driving. However, existing methods suffer from high computational costs and long training time, hindering their deployment on resource-constrained embedded systems. To address this, we propose a fast and robust HAR framework that integrates multidimensional micro-Doppler feature fusion with an edge-optimized Broad Learning System (BLS). Firstly, human activity data are collected using a millimeter-wave radar, and spectral features are extracted through a two-dimensional Fast Fourier Transform (2D-FFT). Subsequently, for the spatial-temporal characteristics of HAR, the pre-activation residual module, multi-scale convolution module, convolutional block attention module and SoftPool are utilized to design a lightweight Hybrid-CNN architecture. Finally, BLS is used for rapid learning and classification, constructing an efficient recognition model. Experimental validation demonstrates three key advantages: Recognition accuracy of 99.02% for nine poses by spectral-convolutional feature fusion; Noise robustness achieving 88.05% accuracy at -5dB SNR via adaptive SoftPool mechanisms; Data efficiency attaining 95.76% accuracy using only 660 training samples through BLS incremental learning. Cross-platform deployment validation demonstrates the method’s efficie...